Recursive Self-Improvement: from Auto Research to Superintelligence — Richard Socher, Recursive
- Source
- youtube.com
- Author
- Latent Space
- Date
Context
Recursive, founded by Richard Socher, aims to use AI to automate AI research itself, an idea called recursive self-improvement. The episode description says the company's early work focuses on AI for AI research, including an AI research system on optimization tasks and work on NVIDIA GPU kernels (small programs that run on graphics processors), where the system found improvements without a team of CUDA experts. The idea is that better research tooling could make later AI development more effective. Episode description
The description reports Socher's claim that the system outperformed humans and their agents on optimization tasks in less than two days. That is a claim about particular early tasks, and no protocol was available. The broader ambition of an "Eureka Machine" that automates invention across science remains a stated goal, not something those results establish.
- Socher says Recursive's system found NVIDIA GPU kernel improvements without relying on a team of CUDA experts.
- He argues AI hard-takeoff scenarios may underestimate physical, compute and economic constraints.
- Socher warns against regulating intelligence itself rather than specific AI applications.
- He says he is less bullish on world models, and the episode also includes his critique of Anthropic's constitution and constitutional AI.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
- LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
Checking sign-in…
Loading comments…






